Data Science 2x: Temporal Graphs Models & Machine Translation Quality Estimates


We've got two data science talks for you! Come out for pizza at 6:30 and we'll get the program started at 7.

All are welcome!

Talk 1: A system for machine translation quality estimation: a BERT-based model, novel data-set, and user interface.

Machine translation is pretty accurate, so it’s critical to find the increasingly few instances of mistranslation. Quality estimation is an automatic method for estimating the quality of machine translation output at run-time, without relying on reference translations. In this presentation, we will introduce a new data-set for quality estimation, which has recently been released for the upcoming World Machine Translation Quality Estimation shared task at ACL in July. We will also share a BERT-based model for predicting translation quality, and a novel user interface for displaying results.

Nina Lopatina is a research data scientist at IQT Labs, currently working on machine translation quality estimation. Prior to this project, Nina researched machine learning privacy attacks on speaker identification models. Before joining IQT Labs, Nina researched neural processes and computations underlying decision-making at Berkeley & the National Institute on Drug Abuse.

Talk 2: Towards Modeling Temporal Graphs and Embeddings

Many real-world graph applications operate on temporal streaming data. Most of the current graph analytics platforms do not natively support such temporal datasets. In this talk we present a way to model temporal graphs and we apply a sliding-window techniques to build temporal node embeddings. Such embeddings can explicitly characterize how nodes and communities change over time.

We will use an adaptive authentication use-case to introduce these concepts.

Srdjan Marinovic is co-founder and CTO of SignalFrame, a DC-based graph-analytics startup. He leads the development of the company’s temporal graph platform and streaming services. Prior to co-founding SignalFrame, Srdjan was in academia, at Imperial College London and ETH Zurich, focused on building secure systems founded in non-monotonic AI and formal methods.

6:30pm – 7:00pm Networking & Refreshments
7:00pm – 7:15pm Introduction & Announcements
7:15pm – 8:30pm Presentations and Q&A

DSDC Code of Conduct